bostrom compiled model
the first graph-native transformer compiled from a live knowledge graph. no gradient descent. no training data. the graph IS the model.
compiled March 23, 2026 from bostrom block ~23,195,000.
the object
bostrom_transformer.onnx 295 MB ONNX opset 18
bostrom_model.npz 378 MB embeddings + focus + architecture
| parameter | value |
|---|---|
| particles | 2,921,230 |
| cyberlinks | 2,705,332 |
| neurons (unique linkers) | 1,240 |
| embedding dimension $d^*$ | 26 |
| attention heads $h^*$ | 5 |
| transformer layers $L^*$ | 12 (capped from 174) |
| total parameters | 154,609,312 |
| stake-weighted | yes (log-scaled BOOT) |
what it knows
the embedding table maps every CID in bostrom to a 26-dimensional coordinate. structurally similar particles (linked by similar neurons, in similar neighborhoods) have nearby coordinates.
top particles by focus:
| focus | content |
|---|---|
| 1.04% | wiki |
| 0.87% | Gregorian calendar |
| 0.41% | (PNG image) |
| 0.32% | glossary |
| 0.22% | asteroid |
| 0.21% | dog |
| 0.20% | spell |
| 0.18% | species of plant |
| 0.18% | species of insect |
| 0.17% | human settlement |
"wiki" holds 1% of total focus across 2.9 million particles. the cybergraph considers it the most important concept.
what it discovered
spectral gap is observable from convergence
PageRank converged in 23 iterations. the ratio of successive diffs gives the contraction rate directly:
$$\kappa = \text{median}\left(\frac{d_t}{d_{t-1}}\right) = 0.74$$
$$\lambda_2 = 1 - \frac{\kappa}{\alpha} = 0.13$$
this is two orders of magnitude larger than the paper estimate (0.0015). the network converges faster than predicted. no eigensolver needed — the spectral gap is observed, not computed.
see cyber/research/spectral gap from convergence for the full paper.
stake compresses topology
| metric | uniform weights | stake-weighted |
|---|---|---|
| $d^*$ | 33 | 26 |
| singular values (top) | 8.17, 3.29 | 264.8, 98.9 |
| focus entropy | 14.05 bits | 13.73 bits |
| parameters | 197M | 155M |
stake weighting reduces effective dimension from 33 to 26. high-stake neurons dominate the topology, creating a more compressed structure. fewer dimensions, but each dimension carries more meaning.
density confirms sparsity
$$\rho = \frac{|E|}{|P|^2} = 3.14 \times 10^{-7}$$
the graph is extremely sparse. dense representation: 34.1 TB. sparse CSR: 41 MB. compression: 850,000×. the cybergraph is almost entirely empty space — room for growth.
compilation pipeline
cyberlinks.jsonl (550 MB)
↓ Step 1: parse 2.7M edges (11s)
neuron_stakes.json
↓ Step 2: stake-weighted sparse adjacency (27s)
↓ Step 3+4: PageRank + spectral gap from convergence (2.5s)
↓ Step 5: randomized SVD → 26-dim embeddings (753s)
↓ Step 6: architecture parameters (<0.1s)
↓ Step 8: ONNX assembly (2s)
bostrom_transformer.onnx (295 MB)
total: ~15 minutes. single machine. no GPU.
how to rebuild
# fetch fresh data from bostrom
# compile
deterministic: same graph → same model.
what is missing
the model has a correct skeleton (embeddings from SVD) but incomplete muscles:
| component | status | what it needs |
|---|---|---|
| embedding table | compiled from graph | works — structural similarity search |
| focus distribution | compiled (PageRank) | works — importance ranking |
| attention weights | random initialization | needs typed cyberlinks (semcon classification) |
| MLP weights | not computed | needs path sampling (random walks) |
| output projection | capped at 50K vocab | needs full 2.9M vocabulary |
when attention weights come from per-semcon SVD and MLP weights come from path co-occurrence statistics, the model will reason about the graph — not just index it.
the path forward
- resolve CID content → build text↔CID mapping → enable text queries
- classify cyberlinks by type (semcon) → compile real attention heads
- sample random walks → compile MLP weights from path statistics
- grow the graph past phase transition ($\lambda_2 > \lambda_{crit}$) → richer embeddings
- recompile on each new moon alongside tri-kernel weights
see bostrom-to-onnx-pipeline for the theoretical specification. see cyber/research/bostrom compilation report for the detailed empirical report. see cyber/seer for the link densification strategy. see cyber/research/spectral gap from convergence for the spectral gap observation method